Image segmentation with boosted multiclass classifiers

Resumen

The goal of image segmentation is to partition an image into different regions, each of them usually endowed with a semantic meaning (i.e. road, house, car, etc). Hence, it is one of the most complex problems faced by computer vision, since in its most general case it entails understanding "what is in the image." In this thesis we address the image segmentation problem from a supervised perspective. Our solution is supported by recentmulti-class Boosting algorithms. Although we approach the problem from a general view point, our algorithms were conceived to address the problem of segmenting electron michrographs (EM) of brain tissue. Most state-of-the-art-approaches for this problem segment one single brain structure at a time. Instead, the tarting point of our proposal is that the simultaneous analysis of several brain structures improves the performance of the segmentation. This lets the algorithm exploit the dependencies among the different classes of structures and find image descriptions that provide better segmentation results, as we prove in our experiments. In this thesis, we develop an image segmentation algorithm that is the state-of-the-art Boosting scheme for segmenting two brain structures, mitochondria and synapses, in EMimages of the brain. To reach this goal we make several contributions in different areas. First, we study the most popular image feature descriptors and compare their performance on EM images, selecting the ones that produce the best segmentation metrics. To reach this result we develop a new feature selection algorithm that is able to select the best set of features and scales to describe a given EM image stack. Secondly, we introduce the Jaccard Curve as a tool to compare the performance of two segmentation algorithms independently of their operation point. We use this curve to compare the performance of our approach with the best algorithms in the literature. As a final contribution of the thesis, and to improve even further the segmentation, we develop a general procedure to optimize typical image segmentation quality indices. To this end we use recent cost-sensitive classification results to learn the costs that optimize the desired index. The experiments show that our procedure, using a pool of simple features, can improve significantly the Jaccard index in the segmentations of synapses and mitochondria as well as that of other natural images.